Ben Thompson declared last week that the AI boom isn't a bubble. Coming from someone who's spent three years carefully hedging that position (even arguing that bubbles can be good), it's a notable shift. And his reasoning, built around the rise of agents, deserves unpacking. Not because the macro argument matters to most of us directly, but because the logic underneath it has real implications for how you think about AI investment right now.

The short version: agents have fundamentally altered the compute equation. They've also altered the competitive equation. And if Thompson's right (and I think he largely is), treating AI as something you'll get to next quarter is now the riskier bet.

Three paradigms, each one hungrier than the last

The piece frames the AI era through three inflection points, each one demanding exponentially more compute.

ChatGPT arrived in November 2022. Brilliant at opening people's eyes, but flawed in two ways that stuck: it hallucinated confidently, and you had to know what to use it for. Inference was cheap. The model spat out an answer and you took what you got.

Reasoning models came next, starting with OpenAI's o1 in September 2024. These models think before they answer, working through problems internally, evaluating whether their response is correct before delivering it. Dramatically more reliable output, but dramatically more compute too. All that internal reasoning generates tokens, and the improved reliability meant people used these models far more heavily.

Then came agents, and this is where the economics get interesting. Claude Code with Opus 4.5, OpenAI's Codex, and their various successors didn't just improve the model. They wrapped it in a harness: software that directs the model, uses deterministic tools to verify results, and retries when things go wrong. The human steps away. The agent handles the loop.

I wrote about this shift when it happened — the moment coding agents crossed from clever assistant to genuine autonomous capability. What I didn't fully appreciate at the time was what it would mean for compute demand. Each agent run doesn't just call a model once. It calls it repeatedly, often through a reasoning model, while also consuming CPU for the harness, the tools, and the verification. Multiply that across an organisation and the numbers get very large very fast. The reliability maths compounds too — each nine of uptime costs as much engineering effort as the last.

This is why every hyperscaler (Amazon, Microsoft, Google, Meta) is reporting that demand for AI compute exceeds supply, even as they collectively spend north of $600 billion on infrastructure in 2026. They're not speculating. They're scrambling to meet orders they've already taken.

Agency without the agents (the human kind)

The part of Thompson's argument that deserves the most attention, at least for smaller businesses, is about agency itself.

Chatbots needed individual humans to take the initiative. You had to decide to use Claude, figure out what to ask, and verify the output yourself. That required agency, the drive and imagination to actually do something with the tool. Most people, it turns out, didn't bother. They used AI the way they used Google: basic questions, tidied-up emails, the occasional summary.

Agents change that equation in a way that matters. Because agents abstract the human away from the model, a single person with initiative can now direct multiple agents working in parallel. You don't need a thousand employees each using a chatbot. You need a handful of people with clear goals, the right agent infrastructure, and the willingness to let the machines run.

He calls this a narrowing of the need for widescale adoption. I'd call it something more immediate: the return of the small, fast team.

I've been banging this drum for a while. AI-native businesses built by small, focused teams can outperform bloated incumbents not by working harder, but by operating at a fundamentally different ratio of people to output. Every article I've written about agentic loops and readiness gaps has circled back to this point. Thompson's analysis puts the structural economics behind it.

If a team of five people can deploy agents that handle the work of fifty, and we're getting closer to that reality every month, then the cost structure of competition shifts permanently. Incumbents don't just face more agile competitors. They face competitors whose economics make the incumbents' entire organisational model untenable.

Why models aren't commodities (and what that means for the value chain)

There's a secondary argument in Thompson's piece that's worth pulling out, because it affects which vendors you should be betting on.

The conventional wisdom, articulated well by Horace Dediu at Asymco, goes something like this: AI models are commoditising fast. DeepSeek built a competitive model for a fraction of what the frontier labs spent. Open-source powers most startups. The smart play is to sit at the integration layer — own the customer relationship and swap models as needed. Apple's licensing of Google's Gemini instead of building its own model is exhibit A.

Thompson argues this logic held during the chatbot era but breaks down with agents. The reason is integration. What made Opus 4.5 suddenly capable wasn't just the model — it was changes to the Claude Code harness that made model and software work together as one system. Agent performance depends on the tight coupling between model and orchestration layer. The whole thing works because the pieces are built together, not snapped in from different vendors.

This matters because profits in technology flow toward integration and away from commoditised, modular parts. If the model-plus-harness combination is where differentiation lives, then Anthropic and OpenAI aren't just AI labs burning cash. They're building the integrated layer where value will accumulate.

Microsoft seems to have reached the same conclusion. After years of positioning itself as model-agnostic, with all that talk about Core AI and interchangeable models, the company's new Copilot Cowork is built specifically with Anthropic's Claude. Integrated, not model-agnostic. And they're charging $99 per seat per month for the E7 bundle that includes it, double the previous top tier.

For businesses choosing AI vendors, the implication is pointed: back the companies building integrated agent systems, not those promising you can plug any model into any harness. The modular story sounds appealing. The integrated products actually work.

The layoff wave is structural, not cyclical

There's a prediction in Thompson's piece that I think will prove accurate and uncomfortable: the coming wave of AI-driven layoffs isn't just companies using AI as cover for correcting pandemic-era over-hiring. Some of it's that, and fair enough. But underneath the convenient timing is a genuine structural shift.

Companies became bloated because that was the only way to scale. You needed humans at every junction: coordinating, communicating, translating intent into action across departments and teams. The coordination costs were enormous, but invisible, because there was no alternative.

Agents are the alternative. Not for everything, not yet, not perfectly. But for enough of those coordination-heavy middle-layer functions that organisations are going to discover their "right size" is significantly smaller than they thought.

The most forward-looking companies won't just trim to a pre-AI baseline. They'll cut deeper, betting that the remaining employees have no choice but to rebuild scale with agents. Because if they don't, a competitor built with agents from the start will do it for them.

I've seen this dynamic playing out already. The readiness gap I wrote about recently — 85% of enterprises wanting agentic AI while 76% can't support it — is partly an infrastructure problem and partly a workforce design problem. Most organisations haven't confronted the question of what their team should look like in an agent-augmented world. They're still thinking about headcount plus AI tools, when the real question is headcount divided by AI capability.

The practical upshot

If Thompson's right that this isn't a bubble — and the structural arguments are strong — then the strategic implications are fairly direct.

The window for cautious, low-commitment AI pilots is closing. The companies building agent workflows now are accumulating compounding advantages that will be difficult to replicate later. You don't need to bet the farm, but you do need to be building real systems, not running demos.

Your team matters more than your tools. Not every employee needs to be an AI power user. But you need people who can direct agents, who understand what to automate, how to verify the output, and when to intervene. Those people are your multiplier.

On the vendor side, the model-agnostic pitch is seductive, but the evidence suggests that tightly integrated agent platforms deliver genuinely better results. The companies building model and harness as one system are the ones worth backing.

And then there's the organisational question that nobody wants to ask out loud. If agents can handle coordination tasks that currently require three people, you don't need three people. That's uncomfortable, but ignoring it doesn't make it less true. It just means a smaller competitor figures it out first.

The bubble question was always the wrong question for most businesses. The right question was always: is this technology going to restructure how work gets done? The answer, increasingly, is yes. And the companies that acted on that answer twelve months ago are already pulling ahead.

The evaluation period is over. The building period started without you.